AI helps to preserve vision

Lenovo company blog, Habr

See the original for references.

The use of artificial intelligence in medicine today allows us to seriously improve the accuracy of diagnosis, make life easier for patients. It is expected that AI will become indispensable in the diagnosis and refinement of diseases. Due to the ability to compare data, collect and synthesize information, the participation of AI in diagnostics should help to qualitatively improve the statistics of medical errors, increase the role of prevention and prevention of diseases.

According to the forecast of the research company Research&Markets, the global artificial intelligence market will grow to $ 5.05 billion by 2020. At the same time, healthcare will become the fastest growing segment. According to international research, the use of artificial intelligence in medicine can increase the profits of companies in the healthcare industry.

In 2016, the share of the European AI market was estimated at $270 million with an expected annual growth of more than 35%. BIS Research estimates that by 2025, the total AI market in healthcare will reach $28 billion with an average annual growth of more than 45.1%, and the AI market for medical imaging and diagnostics will reach $2.5 billion.

AI and the problem of retinal diseases

According to the World Health Organization, vision problems directly affect almost every twentieth inhabitant of the planet, and about 80% of such problems could be avoided with the help of preventive measures. For example, it is very important to detect retinal diseases at an early stage, but ophthalmologists do not have enough resources to thoroughly study and diagnose the disease. Artificial intelligence can help them in this and thereby save the eyesight of millions of patients.

Complications from diabetes (diabetic retinopathy) are one of the leading causes of vision problems. The total number of people with diabetes is expected to double between 2000 and 2030, which will significantly increase the number of cases of eye diseases worldwide.

1

Early diagnosis reduces the number of cases of serious vision loss by more than half. Unfortunately, there is little progress in detecting retinal diseases in the early stages during patient examinations. In the countries that suffer most from these diseases, patients do not undergo regular examinations, and ophthalmologists have a rather low accuracy of correct recognition and diagnosis of retinal diseases with individual in-depth eye examinations. At the same time, unlike other life-threatening diseases that everyone is hearing about today, retinal diseases and visual impairment remain not so noticeable in the eyes of the public. Therefore, the problem is often underestimated.

Under the gaze of artificial intelligence

Artificial intelligence (AI) can potentially contribute to a significant reduction in cases of retinal diseases, helping ophthalmologists to detect the disease more effectively and complementing the human experience. In collaboration with Lenovo, the Barcelona Supercomputing Center (BSC) decided to investigate how AI can improve the accuracy of the screening process and potentially detect retinal disease earlier than it usually happens. AI technology increases the likelihood of early detection of the disease, making the examination of patients more accessible and faster in countries with an underserved population. Moreover, patients can independently undergo an initial examination within a few minutes using their smartphone with a special application.

The future of medicine is disease prevention. 
Therefore, it is important to improve the accuracy of the preliminary diagnosis.

In addition to diabetic retinopathy, eye diseases cause many other pathologies, such as glaucoma, macular degeneration, nevus and epiretinal membrane. Machine learning models make it much easier to identify these various pathologies than current screening methods. Dario Garcia-Gasulla, Honorary researcher at the Barcelona Supercomputing Center, is optimistic about the possibilities of using this technology: "Scaling, training, and validating machine learning models to investigate these vision problems can be a complex process. But the potential is huge, because the same approaches can be applied in other areas of medicine and in many industrial applications."

Model training and overcoming data shortage problems

The problem with training an AI model to detect certain retinal diseases is the lack of "clean" data available for training a neural network. For pathologies with limited data set availability (for example, less than 5000 images), reliable deep learning of a neural network "from scratch" may not be possible. In this case, you can use the "transfer of training".

Automation will give the doctor extra time, which he can use to study the patient's illness and establish the most accurate diagnosis. According to experts from Harvard Medical School, the use of AI technologies will reduce the level of errors in diagnosis by 85%.

The transfer of training is based on models prepared for tasks with larger data sets, which are then reused to solve other tasks with little data availability. Sometimes it is used to isolate signs (extractor). As a result, the transfer of training can also reduce the training time (up to minutes), save hours of research and, ultimately, the costs associated with developing a solution.

Accuracy of detection of various retinal pathologies
with the help of AI, it is 75-91%.

New AI technology

At the International Supercomputing Conference (ISC) in Frankfurt, Lenovo and BSC will show an application demonstrating how learning transfer works. It was created at the Lenovo AI Innovation Center in Morrisville, North Carolina (USA). The application will allow visitors to independently build and train a model through an intuitive interface and thereby play an active role in improving the screening of retinal diseases.

Garcia-Gasulla explains: "The purpose of the demonstration is to show how easy it is to use pre–trained deep neural networks as feature extractors, which become the basis for other, simpler and faster models (in this case SVM). In 10 minutes, each participant will be able to design, train and test the effectiveness of a machine learning model for detecting retinal pathology. Models of conference participants working with the same pathology will be compared and evaluated to find and reward the best model developed during this forum."

LiCO accelerates the training of artificial intelligence models and the traditional deployment of high-performance computing systems by providing an intuitive user interface for managing the software and hardware stack.

Smart medical products, services and processes are already being developed by more than 800 companies, including leading vendors. For such research, Lenovo creates its own AI solutions, including the recently released Lenovo Intelligent Computing Orchestration (LiCO) 5.1 Platform and Lenovo AI Validated Design reference architectures for developing models based on Intel Xeon Scalable and NVIDIA Tesla architectures.

Components of model training. The AI software stack is developing rapidly, new and updated frameworks appear almost monthly. Choosing among the many open source options can take a long time. Lenovo's reference architecture has been tested and configured based on the Lenovo ThinkSystem platform.

AI in medicine: the future has come

The use of artificial intelligence in medicine can revolutionize the healthcare industry through the development of such areas as personalized medicine, diagnostics, development of new drugs, robot-assisted surgery, telemonitoring of chronic diseases, remote patient care, support for making the right medical decisions, identification of medical errors.

Frost & Sullivan agency notes that artificial intelligence technologies increase the accuracy of diagnosis by 30-40%, while the cost of medical care is reduced by half. McKinsey has shown that 36% of functions in medicine can be automated, primarily at the levels of data collection and analysis.

Developments in this direction are actively underway both abroad and in Russia, for example, one of the Russian projects is a system for the diagnosis of diseases, which includes the recognition of pathologies using medical digital images obtained from the results of lung radiography, mammography, computed tomography and ultrasound. The project is an application that can be used on the user's work computer or smartphone. It works on the basis of a neural network trained to recognize pathologies in medical images. The first stage of the project is an analyzer of pathological blood cells and recognition of fundus pathologies. In the future, it will cover such areas as lung radiography, mammography, computed tomography, mobile ultrasound.

And new projects appear almost every year. Many developments are available now. For example, an information and analytical system "SoVgaip‑Analytics" has been launched in Russia to diagnose and form personal therapy for patients with brain diseases. Artificial intelligence in medicine is the future that has already come.

Portal "Eternal youth" http://vechnayamolodost.ru